def action(self, state):
""" Returns an action for a given state.
Public handle to function.
"""
# save state
if self._logging_dir is not None:
policy_id = utils.gen_experiment_id()
self._policy_dir = os.path.join(self._logging_dir,
'policy_output_%s' % (policy_id))
while os.path.exists(self._policy_dir):
policy_id = utils.gen_experiment_id()
self._policy_dir = os.path.join(self._logging_dir,
'policy_output_%s' % (policy_id))
os.mkdir(self._policy_dir)
state_dir = os.path.join(self._policy_dir, 'state')
state.save(state_dir)
# plan action
action = self._action(state)
# save action
if self._logging_dir is not None:
action_dir = os.path.join(self._policy_dir, 'action')
action.save(action_dir)
return action
def benchmark_bin_picking_policy(policy,
# input_dataset_path,
# heap_ids,
# timesteps,
# output_dataset_path,
config,
# excluded_heaps_file):
):
""" Benchmark a bin picking policy.
Parameters
----------
policy : :obj:`Policy`
policy to roll out
input_dataset_path : str
path to the input dataset
heap_ids : list
integer identifiers for the heaps to re-run
timesteps : list
integer timesteps to seed the simulation from
output_dataset_path : str
path to store the results
config : dict
dictionary-like objects containing parameters of the simulator and visualization
"""
# read subconfigs
vis_config = config['vis']
dataset_config = config['dataset']
# read parameters
fully_observed = config['fully_observed']
steps_per_test_case = config['steps_per_test_case']
rollouts_per_garbage_collect = config['rollouts_per_garbage_collect']
debug = config['debug']
im_height = config['state_space']['camera']['im_height']
im_width = config['state_space']['camera']['im_width']
max_obj_per_pile = config['state_space']['object']['max_obj_per_pile']
if debug:
random.seed(SEED)
np.random.seed(SEED)
# read ids
# if len(heap_ids) != len(timesteps):
# raise ValueError('Must provide same number of heap ids and timesteps')
# num_rollouts = len(heap_ids)
num_rollouts = 1
# set dataset params
tensor_config = dataset_config['tensors']
fields_config = tensor_config['fields']
# fields_config['color_ims']['height'] = im_height
# fields_config['color_ims']['width'] = im_width
# fields_config['depth_ims']['height'] = im_height
# fields_config['depth_ims']['width'] = im_width
fields_config['obj_poses']['height'] = POSE_DIM * max_obj_per_pile
fields_config['obj_coms']['height'] = POINT_DIM * max_obj_per_pile
fields_config['obj_ids']['height'] = max_obj_per_pile
fields_config['bin_distances']['height'] = max_obj_per_pile
# matrix has (n choose 2) elements in it
max_distance_matrix_length = int(comb(max_obj_per_pile, 2))
fields_config['distance_matrix']['height'] = max_distance_matrix_length
# sample a process id
proc_id = utils.gen_experiment_id()
# if not os.path.exists(output_dataset_path):
# try:
# os.mkdir(output_dataset_path)
# except:
# logging.warning('Failed to create %s. The dataset path may have been created simultaneously by another process' %(dataset_path))
# proc_id = 'clustering_2'
# output_dataset_path = os.path.join(output_dataset_path, 'dataset_%s' %(proc_id))
# open input dataset
# logging.info('Opening input dataset: %s' % input_dataset_path)
# input_dataset = TensorDataset.open(input_dataset_path)
# open output_dataset
# logging.info('Opening output dataset: %s' % output_dataset_path)
# dataset = TensorDataset(output_dataset_path, tensor_config)
# datapoint = dataset.datapoint_template
# setup logging
# experiment_log_filename = os.path.join(output_dataset_path, 'dataset_generation.log')
# formatter = logging.Formatter('%(asctime)s %(levelname)s: %(message)s')
# hdlr = logging.FileHandler(experiment_log_filename)
# hdlr.setFormatter(formatter)
# logging.getLogger().addHandler(hdlr)
# config.save(os.path.join(output_dataset_path, 'dataset_generation_params.yaml'))
# key mappings
# we add the empty string as a mapping because if you don't evaluate dexnet on the 'before' state of the push
obj_id = 1
obj_ids = {'': 0}
action_ids = {
'ParallelJawGraspAction': 0,
'SuctionGraspAction': 1,
'LinearPushAction': 2
}
# add action ids
reverse_action_ids = utils.reverse_dictionary(action_ids)
# dataset.add_metadata('action_ids', reverse_action_ids)
# perform rollouts
n = 0
rollout_start = time.time()
current_heap_id = None
while n < num_rollouts:
# create env
create_start = time.time()
bin_picking_env = GraspingEnv(config, vis_config)
create_stop = time.time()
logging.info('Creating env took %.3f sec' %(create_stop-create_start))
# perform rollouts
rollouts_remaining = num_rollouts - n
for i in range(min(rollouts_per_garbage_collect, rollouts_remaining)):
# log current rollout
logging.info('\n')
if n % vis_config['log_rate'] == 0:
logging.info('Rollout: %03d' %(n))
try:
# mark rollout status
data_saved = False
num_steps = 0
# read heap id
# heap_id = heap_ids[n]
# timestep = timesteps[n]
# while heap_id == current_heap_id:# or heap_id < 81:#[226, 287, 325, 453, 469, 577, 601, 894, 921]: 26
# n += 1
# heap_id = heap_ids[n]
# timestep = timesteps[n]
push_logger = logging.getLogger('push')
# push_logger.info('~')
# push_logger.info('Heap ID %d' % heap_id)
# current_heap_id = heap_id
# reset env
reset_start = time.time()
# bin_picking_env.reset_from_dataset(input_dataset,
# heap_id,
# timestep)
bin_picking_env.reset()
state = bin_picking_env.state
environment = bin_picking_env.environment
if fully_observed:
observation = None
else:
observation = bin_picking_env.observation
policy.set_environment(environment)
reset_stop = time.time()
# add objects to mapping
for obj_key in state.obj_keys:
if obj_key not in obj_ids.keys():
obj_ids[obj_key] = obj_id
obj_id += 1
push_logger.info(obj_key)
# save id mappings
reverse_obj_ids = utils.reverse_dictionary(obj_ids)
# dataset.add_metadata('obj_ids', reverse_obj_ids)
# store datapoint env params
# datapoint['heap_ids'] = current_heap_id
# datapoint['camera_poses'] = environment.camera.T_camera_world.vec
# datapoint['camera_intrs'] = environment.camera.intrinsics.vec
# datapoint['robot_poses'] = environment.robot.T_robot_world.vec
# render
if vis_config['initial_state']:
vis3d.figure()
bin_picking_env.render_3d_scene()
vis3d.pose(environment.robot.T_robot_world)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# observe
if vis_config['initial_obs']:
vis2d.figure()
vis2d.imshow(observation, auto_subplot=True)
vis2d.show()
# rollout on current satte
done = False
failed = False
# if isinstance(policy, SingulationFullRolloutPolicy):
# policy.reset_num_failed_grasps()
while not done:
if vis_config['step_stats']:
logging.info('Heap ID: %s' % heap_id)
logging.info('Timestep: %s' % bin_picking_env.timestep)
# get action
policy_start = time.time()
if fully_observed:
action = policy.action(state)
else:
action = policy.action(observation)
policy_stop = time.time()
logging.info('Composite Policy took %.3f sec' %(policy_stop-policy_start))
# render scene before
if vis_config['action']:
#gripper = bin_picking_env.gripper(action)
vis3d.figure()
# GRASPINGENV
# bin_picking_env.render_3d_scene(render_camera=False, workspace_objs_wireframe=False)
bin_picking_env.render_3d_scene()
if isinstance(action, GraspAction):
vis3d.gripper(gripper, action.grasp(gripper))
#if isinstance(action, LinearPushAction):
else:
# # T_start_world = action.T_begin_world * gripper.T_mesh_grasp
# # T_end_world = action.T_end_world * gripper.T_mesh_grasp
# #start_point = action.T_begin_world.translation
# start_point = action['start']
# #end_point = action.T_end_world.translation
# end_point = action['end']
# vec = (end_point - start_point) / np.linalg.norm(end_point-start_point) if np.linalg.norm(end_point-start_point) > 0 else end_point-start_point
# #h1 = np.array([[0.7071,-0.7071,0],[0.7071,0.7071,0],[0,0,1]]).dot(vec)
# #h2 = np.array([[0.7071,0.7071,0],[-0.7071,0.7071,0],[0,0,1]]).dot(vec)
# arrow_len = np.linalg.norm(start_point - end_point)
# h1 = (end_point - start_point + np.array([0,0,arrow_len])) / (arrow_len*math.sqrt(2))
# h2 = (end_point - start_point - np.array([0,0,arrow_len])) / (arrow_len*math.sqrt(2))
# shaft_points = [start_point, end_point]
# head_points = [end_point - 0.03*h2, end_point, end_point - 0.03*h1]
# #vis3d.plot3d(shaft_points, color=[0,0,1])
# #vis3d.plot3d(head_points, color=[0,0,1])
# Displaying all potential topple points
for vertex, prob in zip(action['vertices'], action['probabilities']):
color = np.array([min(1, 2*(1-prob)), min(2*prob, 1), 0])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
for vertex in action['bottom_points']:
color = np.array([0,0,1])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
vis3d.points(Point(action['com'], 'world'), scale=.005, color=np.array([0,0,1]))
vis3d.points(Point(np.array([0,0,0]), 'world'), scale=.005, color=np.array([0,1,0]))
#set_of_lines = action['set_of_lines']
#for i, line in enumerate(set_of_lines):
# color = str(bin(i+1))[2:].zfill(3)
# color = np.array([color[2], color[1], color[0]])
# vis3d.plot3d(line, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# Show
vis3d.figure()
bin_picking_env.render_3d_scene()
final_pose_ind = action['final_pose_ind'] / np.amax(action['final_pose_ind'])
for vertex, final_pose_ind in zip(action['vertices'], final_pose_ind):
color = np.array([0, min(1, 2*(1-prob)), min(2*prob, 1)])
vis3d.points(Point(vertex, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
color=np.array([0,0,1])
original_pose = state.obj.T_obj_world
pose_num = 0
for pose, edge_point1, edge_point2 in zip(action['final_poses'], action['bottom_points'], np.roll(action['bottom_points'],-1,axis=0)):
print 'Pose:', pose_num
pose_num += 1
pose = pose.T_obj_table
vis3d.figure()
state.obj.T_obj_world = original_pose
bin_picking_env.render_3d_scene()
vis3d.points(Point(edge_point1, 'world'), scale=.0005, color=color)
vis3d.points(Point(edge_point2, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
vis3d.figure()
state.obj.T_obj_world = pose
bin_picking_env.render_3d_scene()
vis3d.points(Point(edge_point1, 'world'), scale=.0005, color=color)
vis3d.points(Point(edge_point2, 'world'), scale=.0005, color=color)
vis3d.show(starting_camera_pose=CAMERA_POSE)
#vis3d.save('/home/mjd3/Pictures/weird_pics/%d_%d_before.png' % (heap_id, bin_picking_env.timestep), starting_camera_pose=CAMERA_POSE)
# store datapoint pre-step data
j = 0
obj_poses = np.zeros(fields_config['obj_poses']['height'])
obj_coms = np.zeros(fields_config['obj_coms']['height'])
obj_ids_vec = np.iinfo(np.uint32).max * np.ones(fields_config['obj_ids']['height'])
for obj_state in state.obj_states:
obj_poses[j*POSE_DIM:(j+1)*POSE_DIM] = obj_state.T_obj_world.vec
obj_coms[j*POINT_DIM:(j+1)*POINT_DIM] = obj_state.center_of_mass
obj_ids_vec[j] = obj_ids[obj_state.key]
j += 1
action_poses = np.zeros(fields_config['action_poses']['height'])
#if isinstance(action, GraspAction):
# action_poses[:7] = action.T_grasp_world.vec
#else:
# action_poses[:7] = action.T_begin_world.vec
# action_poses[7:] = action.T_end_world.vec
# if isinstance(policy, SingulationMetricsCompositePolicy):
# actual_distance_matrix_length = int(comb(len(state.objs), 2))
# bin_distances = np.append(action.metadata['bin_distances'],
# np.zeros(max_obj_per_pile-len(state.objs))
# )
# distance_matrix = np.append(action.metadata['distance_matrix'],
# np.zeros(max_distance_matrix_length - actual_distance_matrix_length)
# )
# datapoint['bin_distances'] = bin_distances
# datapoint['distance_matrix'] = distance_matrix
# datapoint['T_begin_world'] = action.T_begin_world.matrix
# datapoint['T_end_world'] = action.T_end_world.matrix
# datapoint['parallel_jaw_best_q_value'] = action.metadata['parallel_jaw_best_q_value']
# # datapoint['parallel_jaw_mean_q_value'] = action.metadata['parallel_jaw_mean_q_value']
# # datapoint['parallel_jaw_num_grasps'] = action.metadata['parallel_jaw_num_grasps']
# datapoint['suction_best_q_value'] = action.metadata['suction_best_q_value']
# # datapoint['suction_mean_q_value'] = action.metadata['suction_mean_q_value']
# # datapoint['suction_num_grasps'] = action.metadata['suction_num_grasps']
# # logging.info('Suction Q: %f, PJ Q: %f' % (action.metadata['suction_q_value'], action.metadata['parallel_jaw_q_value']))
# # datapoint['obj_index'] = action.metadata['obj_index']
# # datapoint['parallel_jaw_best_q_value_single'] = action.metadata['parallel_jaw_best_q_value_single']
# # datapoint['suction_best_q_value_single'] = action.metadata['suction_best_q_value_single']
# datapoint['singulated_obj_index'] = action.metadata['singulated_obj_index']
# datapoint['parallel_jaw_grasped_obj_index'] = obj_ids[action.metadata['parallel_jaw_grasped_obj_key']]
# datapoint['suction_grasped_obj_index'] = obj_ids[action.metadata['suction_grasped_obj_key']]
# else:
# datapoint['bin_distances'] = np.zeros(max_obj_per_pile)
# datapoint['distance_matrix'] = np.zeros(max_distance_matrix_length)
# datapoint['T_begin_world'] = np.zeros((4,4))
# datapoint['T_end_world'] = np.zeros((4,4))
# datapoint['parallel_jaw_best_q_value'] = -1
# datapoint['suction_best_q_value'] = -1
# datapoint['singulated_obj_index'] = -1
# datapoint['parallel_jaw_grasped_obj_index'] = -1
# datapoint['suction_grasped_obj_index'] = -1
# policy_id = 0
# if 'policy_id' in action.metadata.keys():
# policy_id = action.metadata['policy_id']
# greedy_q_value = 0
# if 'greedy_q_value' in action.metadata.keys():
# greedy_q_value = action.metadata['greedy_q_value']
# datapoint['timesteps'] = bin_picking_env.timestep
# datapoint['obj_poses'] = obj_poses
# datapoint['obj_coms'] = obj_coms
# datapoint['obj_ids'] = obj_ids_vec
# # if bin_picking_env.render_mode == RenderMode.RGBD:
# # color_data = observation.color.raw_data
# # depth_data = observation.depth.raw_data
# # elif bin_picking_env.render_mode == RenderMode.DEPTH:
# # color_data = np.zeros(observation.shape).astype(np.uint8)
# # depth_data = observation.raw_data
# # elif bin_picking_env.render_mode == RenderMode.COLOR:
# # color_data = observation.raw_data
# # depth_data = np.zeros(observation.shape)
# # datapoint['color_ims'] = color_data
# # datapoint['depth_ims'] = depth_data
# datapoint['action_ids'] = action_ids[type(action).__name__]
# datapoint['action_poses'] = action_poses
# datapoint['policy_ids'] = policy_id
# datapoint['greedy_q_values'] = greedy_q_value
# datapoint['pred_q_values'] = action.q_value
# step the policy
#observation, reward, done, info = bin_picking_env.step(action)
#state = bin_picking_env.state
state.objs[0].T_obj_world = action['final_state']
# if isinstance(policy, SingulationFullRolloutPolicy):
# policy.grasp_succeeds(info['grasp_succeeds'])
# debugging info
if vis_config['step_stats']:
logging.info('Action type: %s' %(type(action).__name__))
logging.info('Action Q-value: %.3f' %(action.q_value))
logging.info('Reward: %d' %(reward))
logging.info('Policy took %.3f sec' %(policy_stop-policy_start))
logging.info('Num objects remaining: %d' %(bin_picking_env.num_objects))
if info['cleared_pile']:
logging.info('Cleared pile!')
# # store datapoint post-step data
# datapoint['rewards'] = reward
# datapoint['grasp_metrics'] = info['grasp_metric']
# datapoint['collisions'] = 1 * info['collides']
# datapoint['collisions_with_env'] = 1 * info['collides_with_static_obstacles']
# datapoint['grasped_obj_ids'] = obj_ids[info['grasped_obj_key']]
# datapoint['cleared_pile'] = 1 * info['cleared_pile']
# # store datapoint
# # dataset.add(datapoint)
# data_saved = True
# render observation
if vis_config['obs']:
vis2d.figure()
vis2d.imshow(observation, auto_subplot=True)
vis2d.show()
# render scene after
if vis_config['state']:
vis3d.figure()
bin_picking_env.render_3d_scene(render_camera=False)
vis3d.show(starting_camera_pose=CAMERA_POSE)
# vis3d.save('/home/mjd3/Pictures/weird_pics/%d_%d_after.png' % (heap_id, bin_picking_env.timestep), starting_camera_pose=CAMERA_POSE)
state.objs[0].T_obj_world = action['tmpR']
vis3d.figure()
bin_picking_env.render_3d_scene(render_camera=False)
vis3d.show(starting_camera_pose=CAMERA_POSE)
state.objs[0].T_obj_world = action['final_state']
# increment the number of steps
num_steps += 1
if num_steps >= steps_per_test_case:
done = True
except NoActionFoundException as e:
logging.warning('The policy failed to plan an action!')
done = True
except Exception as e:
# log an error
logging.warning('Rollout failed!')
logging.warning('%s' %(str(e)))
logging.warning(traceback.print_exc())
# if debug:
# raise
# reset env
del bin_picking_env
gc.collect()
bin_picking_env = BinPickingEnv(config, vis_config)
# terminate current rollout
failed = True
done = True
# update test case id
n += 1
# dataset.flush()
# logging.info("\n\nflushing")
# logging.info("exiting")
# sys.exit()
# garbage collect
del bin_picking_env
gc.collect()
# return the dataset
# dataset.flush()
# log time
rollout_stop = time.time()
logging.info('Rollouts took %.3f sec' %(rollout_stop-rollout_start))
return dataset
def test_single_read_write(self):
# seed
np.random.seed(SEED)
random.seed(SEED)
# open dataset
create_successful = True
try:
dataset = TensorDataset(TEST_TENSOR_DATASET_NAME, TENSOR_CONFIG)
except:
create_successful = False
self.assertTrue(create_successful)
# check field names
write_datapoint = dataset.datapoint_template
for field_name in write_datapoint.keys():
self.assertTrue(field_name in dataset.field_names)
# add the datapoint
write_datapoint['float_value'] = np.random.rand()
write_datapoint['int_value'] = int(100 * np.random.rand())
write_datapoint['str_value'] = utils.gen_experiment_id()
write_datapoint['vector_value'] = np.random.rand(HEIGHT)
write_datapoint['matrix_value'] = np.random.rand(HEIGHT, WIDTH)
write_datapoint['image_value'] = np.random.rand(
HEIGHT, WIDTH, CHANNELS)
dataset.add(write_datapoint)
# check num datapoints
self.assertTrue(dataset.num_datapoints == 1)
# add metadata
metadata_num = np.random.rand()
dataset.add_metadata('test', metadata_num)
# check written arrays
dataset.flush()
for field_name in dataset.field_names:
filename = os.path.join(TEST_TENSOR_DATASET_NAME, 'tensors',
'%s_00000.npz' % (field_name))
value = np.load(filename)['arr_0']
if isinstance(value[0], str):
self.assertTrue(value[0] == write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(value[0], write_datapoint[field_name]))
# re-open the dataset
del dataset
dataset = TensorDataset.open(TEST_TENSOR_DATASET_NAME)
# read metadata
self.assertTrue(np.allclose(dataset.metadata['test'], metadata_num))
# read datapoint
read_datapoint = dataset.datapoint(0)
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(
read_datapoint[field_name] == write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# check iterator
for read_datapoint in dataset:
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# read individual fields
for field_name in dataset.field_names:
read_datapoint = dataset.datapoint(0, field_names=[field_name])
if isinstance(read_datapoint[field_name], str):
self.assertTrue(
read_datapoint[field_name] == write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# re-open the dataset in write-only
del dataset
dataset = TensorDataset.open(TEST_TENSOR_DATASET_NAME,
access_mode=READ_WRITE_ACCESS)
# delete datapoint
dataset.delete_last()
# check that the dataset is correct
self.assertTrue(dataset.num_datapoints == 0)
self.assertTrue(dataset.num_tensors == 0)
for field_name in dataset.field_names:
filename = os.path.join(TEST_TENSOR_DATASET_NAME, 'tensors',
'%s_00000.npz' % (field_name))
self.assertFalse(os.path.exists(filename))
# remove dataset
if os.path.exists(TEST_TENSOR_DATASET_NAME):
shutil.rmtree(TEST_TENSOR_DATASET_NAME)
def finetune_classification_cnn(config):
""" Main function. """
# read params
dataset = config['dataset']
x_names = config['x_names']
y_name = config['y_name']
model_dir = config['model_dir']
debug = config['debug']
num_classes = None
if 'num_classes' in config.keys():
num_classes = config['num_classes']
batch_size = config['training']['batch_size']
train_pct = config['training']['train_pct']
model_save_period = config['training']['model_save_period']
data_aug_config = config['data_augmentation']
preproc_config = config['preprocessing']
iterator_config = config['data_iteration']
model_config = config['model']
base_model_config = model_config['base']
optimization_config = config['optimization']
train_config = config['training']
generator_image_shape = None
if 'image_shape' in data_aug_config.keys():
generator_image_shape = data_aug_config['image_shape']
optimizer_name = optimization_config['optimizer']
model_params = {}
if 'params' in model_config.keys():
model_params = model_config['params']
base_model_params = {}
if 'params' in base_model_config.keys():
base_model_params = base_model_config['params']
if debug:
seed = 108
random.seed(seed)
np.random.seed(seed)
# generate model dir
if not os.path.exists(model_dir):
os.mkdir(model_dir)
model_id = utils.gen_experiment_id()
model_dir = os.path.join(model_dir, 'model_%s' % (model_id))
if not os.path.exists(model_dir):
os.mkdir(model_dir)
logging.info('Saving model to %s' % (model_dir))
latest_model_filename = os.path.join(model_dir, 'weights_{epoch:05d}.h5')
best_model_filename = os.path.join(model_dir, 'weights.h5')
# save config
training_config_filename = os.path.join(model_dir, 'training_config.yaml')
config.save(training_config_filename)
# open dataset
dataset = TensorDataset.open(dataset)
# split dataset
indices_filename = os.path.join(model_dir, 'splits.npz')
if os.path.exists(indices_filename):
indices = np.load(indices_filename)['arr_0'].tolist()
train_indices = indices['train']
val_indices = indices['val']
else:
train_indices, val_indices = dataset.split(train_pct)
indices = np.array({'train': train_indices, 'val': val_indices})
np.savez_compressed(indices_filename, indices)
num_train = train_indices.shape[0]
num_val = val_indices.shape[0]
val_steps = int(np.ceil(float(num_val) / batch_size))
# init generator
train_generator_filename = os.path.join(model_dir,
'train_preprocessor.pkl')
val_generator_filename = os.path.join(model_dir, 'val_preprocessor.pkl')
if os.path.exists(train_generator_filename):
logging.info('Loading generators')
train_generator = pkl.load(open(train_generator_filename, 'rb'))
val_generator = pkl.load(open(val_generator_filename, 'rb'))
else:
logging.info('Fitting generator')
train_generator = TensorDataGenerator(num_classes=num_classes,
**data_aug_config)
val_generator = TensorDataGenerator(
featurewise_center=data_aug_config['featurewise_center'],
featurewise_std_normalization=data_aug_config[
'featurewise_std_normalization'],
image_shape=generator_image_shape,
num_classes=num_classes)
fit_start = time.time()
train_generator.fit(dataset,
x_names,
y_name,
indices=train_indices,
**preproc_config)
val_generator.mean = train_generator.mean
val_generator.std = train_generator.std
val_generator.min_output = train_generator.min_output
val_generator.max_output = train_generator.max_output
val_generator.num_classes = train_generator.num_classes
fit_stop = time.time()
logging.info('Generator fit took %.3f sec' % (fit_stop - fit_start))
pkl.dump(train_generator, open(train_generator_filename, 'wb'))
pkl.dump(val_generator, open(val_generator_filename, 'wb'))
if num_classes is None:
num_classes = int(train_generator.num_classes)
# init iterator
train_iterator = train_generator.flow_from_dataset(dataset,
x_names,
y_name,
indices=train_indices,
batch_size=batch_size,
**iterator_config)
val_iterator = val_generator.flow_from_dataset(dataset,
x_names,
y_name,
indices=val_indices,
batch_size=batch_size,
**iterator_config)
# setup model
base_cnn = ClassificationCNN.open(base_model_config['model'],
base_model_config['type'],
input_name=x_names[0],
**base_model_params)
cnn = FinetunedClassificationCNN(base_cnn=base_cnn,
name='dexresnet',
num_classes=num_classes,
output_name=y_name,
im_preprocessor=val_generator,
**model_params)
# setup training
cnn.freeze_base_cnn()
if optimizer_name == 'sgd':
optimizer = SGD(lr=optimization_config['lr'],
momentum=optimization_config['momentum'])
elif optimizer_name == 'adam':
optimizer = Adam(lr=optimization_config['lr'])
else:
raise ValueError('Optimizer %s not supported!' % (optimizer_name))
model = cnn.model
model.compile(optimizer=optimizer,
loss=optimization_config['loss'],
metrics=optimization_config['metrics'])
# train
steps_per_epoch = int(np.ceil(float(num_train) / batch_size))
latest_model_ckpt = ModelCheckpoint(latest_model_filename,
period=model_save_period)
best_model_ckpt = ModelCheckpoint(best_model_filename,
save_best_only=True,
period=model_save_period)
train_history_cb = TrainHistory(model_dir)
callbacks = [latest_model_ckpt, best_model_ckpt, train_history_cb]
history = model.fit_generator(
train_iterator,
steps_per_epoch=steps_per_epoch,
epochs=train_config['epochs'],
callbacks=callbacks,
validation_data=val_iterator,
validation_steps=val_steps,
class_weight=train_config['class_weight'],
use_multiprocessing=train_config['use_multiprocessing'])
# save model
cnn.save(model_dir)
# save history
history_filename = os.path.join(model_dir, 'history.pkl')
pkl.dump(history.history, open(history_filename, 'wb'))
def test_multi_tensor_read_write(self):
# seed
np.random.seed(SEED)
random.seed(SEED)
# open dataset
dataset = TensorDataset(TEST_TENSOR_DATASET_NAME, TENSOR_CONFIG)
write_datapoints = []
for i in range(DATAPOINTS_PER_FILE + 1):
write_datapoint = {}
write_datapoint['float_value'] = np.random.rand()
write_datapoint['int_value'] = int(100 * np.random.rand())
write_datapoint['str_value'] = utils.gen_experiment_id()
write_datapoint['vector_value'] = np.random.rand(HEIGHT)
write_datapoint['matrix_value'] = np.random.rand(HEIGHT, WIDTH)
write_datapoint['image_value'] = np.random.rand(
HEIGHT, WIDTH, CHANNELS)
dataset.add(write_datapoint)
write_datapoints.append(write_datapoint)
# check num datapoints
self.assertTrue(dataset.num_datapoints == DATAPOINTS_PER_FILE + 1)
self.assertTrue(dataset.num_tensors == 2)
# check read
dataset.flush()
del dataset
dataset = TensorDataset.open(TEST_TENSOR_DATASET_NAME,
access_mode=READ_WRITE_ACCESS)
for i, read_datapoint in enumerate(dataset):
write_datapoint = write_datapoints[i]
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
for i, read_datapoint in enumerate(dataset):
# check iterator item
write_datapoint = write_datapoints[i]
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# check random item
ind = np.random.choice(dataset.num_datapoints)
write_datapoint = write_datapoints[ind]
read_datapoint = dataset.datapoint(ind)
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# check deletion
dataset.delete_last()
self.assertTrue(dataset.num_datapoints == DATAPOINTS_PER_FILE)
self.assertTrue(dataset.num_tensors == 1)
for field_name in dataset.field_names:
filename = os.path.join(TEST_TENSOR_DATASET_NAME, 'tensors',
'%s_00001.npz' % (field_name))
dataset.add(write_datapoints[-1])
for write_datapoint in write_datapoints:
dataset.add(write_datapoint)
self.assertTrue(dataset.num_datapoints == 2 *
(DATAPOINTS_PER_FILE + 1))
self.assertTrue(dataset.num_tensors == 3)
# check valid
for i in range(dataset.num_datapoints):
read_datapoint = dataset.datapoint(i)
write_datapoint = write_datapoints[i % (len(write_datapoints))]
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# check read then write out of order
ind = np.random.choice(DATAPOINTS_PER_FILE)
write_datapoint = write_datapoints[ind]
read_datapoint = dataset.datapoint(ind)
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(
read_datapoint[field_name] == write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
write_datapoint = write_datapoints[0]
dataset.add(write_datapoint)
read_datapoint = dataset.datapoint(dataset.num_datapoints - 1)
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(
read_datapoint[field_name] == write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
dataset.delete_last()
# check data integrity
for i, read_datapoint in enumerate(dataset):
write_datapoint = write_datapoints[i % len(write_datapoints)]
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# delete last
dataset.delete_last(len(write_datapoints))
self.assertTrue(dataset.num_datapoints == DATAPOINTS_PER_FILE + 1)
self.assertTrue(dataset.num_tensors == 2)
for i, read_datapoint in enumerate(dataset):
write_datapoint = write_datapoints[i]
for field_name in dataset.field_names:
if isinstance(read_datapoint[field_name], str):
self.assertTrue(read_datapoint[field_name] ==
write_datapoint[field_name])
else:
self.assertTrue(
np.allclose(read_datapoint[field_name],
write_datapoint[field_name]))
# remove dataset
if os.path.exists(TEST_TENSOR_DATASET_NAME):
shutil.rmtree(TEST_TENSOR_DATASET_NAME)